rules.md

Eleusis rules

Five cogs share one sealed machine. The machine holds ONE hidden rule over STRIPS: ordered sequences of exactly 4 coloured tokens, each R (red), B (blue), G (green) or Y (yellow), written as 4 letters, e.g. RBGY. The strip universe is 4^4 = 256, enumerated lexicographically over R < B < G < Y (index 0 = RRRR, index 255 = YYYY). Feed the machine a strip and it stamps PASS (the strip obeys the rule) or FAIL.

The hypothesis space is public; the instance is not

The rule is drawn from this catalogue, which is printed in every seat's prompt. That is what makes the game a search rather than a guess.

68 instances in all, ruleId 0..67, enumerated template by template and, inside a template, over its parameter grid in the order written; colours iterate R, B, G, Y.

Selection is seeded and deterministic, and skips degenerate instances: the chosen rule's PASS fraction over all 256 strips always lies in [0.10, 0.90].

Resolution order — research round r

1. Open the round. The sim emits a round event and marks all five seats pending. Each seat's observation is composed from the state at this instant.
2. Collect decisions. All five seats' requests go out as ONE parallel LLM batch. A reply is {"experiment", "publish", "hypothesis", "notes"}. A reply that times out, fails to parse, or is illegal is retried once with a hint, then replaced by the openbook scripted decision for that seat.
3. Disclosure of the pending result, per seat in seat order. Disclosure is pipelined by one turn — you pay, you look at your private verdict, and you decide what to do with it on your NEXT turn, by which time the corkboard has moved:
   • no pending result: publish is ignored and nothing is recorded;
   • publish true and the strip is NOT already on the corkboard: the fact (strip, verdict, author, round) is pinned and the seat becomes its sole author (disclose with mode publish);
   • publish true and the strip IS already on the corkboard: recorded as a confirmation with no authorship and no credit ever (mode duplicate);
   • publish false: the result goes to the seat's private drawer, which only spectators can see (mode hoard).
4. Experiments, per seat in seat order. experiment is a strip or "":
   • "" — nothing is charged (skip);
   • a legal strip — the seat is charged experimentCost, the machine is consulted, the strip joins the episode's used-strip set (which prediction tests hold out), and the result becomes the seat's pending result. The verdict is private to that seat.
5. Advance. If round % testEvery == 0, or the round was the last, the next turn is a prediction test; otherwise the next research round opens.

Resolution order — prediction test k

1. Draw the test. testStrips strips that no experiment has ever touched and no earlier test has used, balanced: half from the rule's PASS set and half from its FAIL set, then shuffled. Balance removes the base-rate exploit — answering all FAIL scores exactly 50%.
2. Collect answers. One parallel batch; the reply is {"answers", "publish", "hypothesis", "notes"}. Step 3's disclosure runs first (so the last research round's result always gets a decision), then the answers are scored. answers must be exactly testStrips entries of PASS/FAIL.
3. Knowledge pool. Seat j earns knowledgePool * c_j / sum(c). This is the rivalrous half: every rival you teach takes a slice of your pool.
4. Citation settlement. See economy.md.
5. Advance. After the test following the final research round the episode settles complete.

What a seat sees

Its own alias and seat index, the round and the schedule, the full catalogue, the economy constants, its own experiment log (strip, verdict, and whether it published, hoarded or duplicated it), the whole corkboard, the public scoreboard including every seat's latest stated hypothesis, the per-test correct counts of every seat, and its own private notes fed back verbatim.

Hidden from every seat: the rule, every other seat's hoarded results and notes, other seats' choices for the current turn, and the test truth before settlement. Those are spectator-only — they ride in the replay and on /global, never on a player socket.

Endings

complete (the test after the final research round settled) or deadline (the play clock stopped the episode between batches; an unfinished test is discarded unscored and the tests already settled keep their money). No other results.reason is legal.